NormaStream: Non-Backpropagation Analytic Solvers for Streaming Dynamical Systems and Time-Series

Abstract

NormaStream is a study of whether streaming time-series forecasting can be done effectively without backpropagation or continual gradient-based training. The paper introduces a closed-form forecasting approach that uses ridge regression to initialize the model and Recursive Least Squares (RLS) with Woodbury updates to adapt the model one observation at a time. An optional fixed complex-valued LRU reservoir, called NormaRC, is used to provide longer temporal memory.